{ "base_url": "http://127.0.0.1:8001/v1", "clock_domain_id": null, "completed": 3, "concurrency": 1, "configuration_fingerprint": null, "context_distribution": { "380000": 3 }, "context_seed": 0, "context_tokens": 380000, "controlled_output_policy": "observe", "engine": "sglang-glm53-sm120-v0.1.1-rc14", "failed": 0, "failures": [], "finished_at": "2026-09-08T05:23:29Z", "finished_at_unix_ns": 1788845009719808105, "gpu": "2x RTX PRO 6000 Blackwell Max-Q TP2 Linux", "manifest": null, "max_context_tokens": 393216, "max_tokens": 256, "measurement_protocol": "capacity-v3", "metric_population": { "aggregate_throughput_eligible": true, "attempted_requests": 3, "excluded_requests": 0, "failure_policy": "failed or contract-invalid responses retain sanitized timing and metadata but are excluded from performance distributions", "p99_interpretation": "descriptive-nearest-rank-only", "p99_interpretation_by_metric": { "decode_tok_s": "descriptive-nearest-rank-only", "e2e": "descriptive-nearest-rank-only", "effective_prefill_tok_s": "descriptive-nearest-rank-only", "tpot": "descriptive-nearest-rank-only", "ttft": "descriptive-nearest-rank-only" }, "p99_minimum_recommended_samples": 100, "p99_samples_by_metric": { "decode_tok_s": 3, "e2e": 3, "effective_prefill_tok_s": 3, "tpot": 3, "ttft": 3 }, "performance_eligible_requests": 3, "successful_requests": 3 }, "metrics": { "completion_tokens_max": 51.0, "completion_tokens_mean": 44.333333333333336, "completion_tokens_mean_ci95_high": 54.601272648394456, "completion_tokens_mean_ci95_low": 34.065394018272215, "completion_tokens_min": 34.0, "completion_tokens_p25": 34.0, "completion_tokens_p50": 48.0, "completion_tokens_p75": 51.0, "completion_tokens_p90": 51.0, "completion_tokens_p95": 51.0, "completion_tokens_p99": 51.0, "completion_tokens_samples": 3, "completion_tokens_stddev": 9.073771725877465, "content_chunks": 57, "content_chunks_s": null, "controlled_output_exact_adherence": 0, "controlled_output_exact_adherence_rate": null, "controlled_output_samples": 0, "controlled_output_unobservable": 0, "decode_tok_s_max": 138.46226757479639, "decode_tok_s_mean": 126.2620889220376, "decode_tok_s_mean_ci95_high": 138.21908926600358, "decode_tok_s_mean_ci95_low": 114.30508857807163, "decode_tok_s_min": 120.03785732056966, "decode_tok_s_p25": 120.03785732056966, "decode_tok_s_p50": 120.28614187074675, "decode_tok_s_p75": 138.46226757479639, "decode_tok_s_p90": 138.46226757479639, "decode_tok_s_p95": 138.46226757479639, "decode_tok_s_p99": 138.46226757479639, "decode_tok_s_samples": 3, "decode_tok_s_stddev": 10.566393929524283, "e2e_max_ms": 62054.61933100014, "e2e_mean_ci95_high_ms": 61996.78331199598, "e2e_mean_ci95_low_ms": 54173.62592533938, "e2e_mean_ms": 58085.20461866768, "e2e_min_ms": 55738.134884999454, "e2e_p25_ms": 55738.134884999454, "e2e_p50_ms": 56462.859640003444, "e2e_p75_ms": 62054.61933100014, "e2e_p90_ms": 62054.61933100014, "e2e_p95_ms": 62054.61933100014, "e2e_p99_ms": 62054.61933100014, "e2e_samples": 3, "e2e_stddev_ms": 3456.659711555354, "effective_prefill_tok_s_max": 5503.949088414699, "effective_prefill_tok_s_mean": 5285.990335040077, "effective_prefill_tok_s_mean_ci95_high": 5638.988180281117, "effective_prefill_tok_s_mean_ci95_low": 4932.992489799038, "effective_prefill_tok_s_min": 4928.657311905739, "effective_prefill_tok_s_p25": 4928.657311905739, "effective_prefill_tok_s_p50": 5425.364604799794, "effective_prefill_tok_s_p75": 5503.949088414699, "effective_prefill_tok_s_p90": 5503.949088414699, "effective_prefill_tok_s_p95": 5503.949088414699, "effective_prefill_tok_s_p99": 5503.949088414699, "effective_prefill_tok_s_samples": 3, "effective_prefill_tok_s_stddev": 311.9439810815386, "generation_max_ms": 415.6754819996422, "generation_mean_ci95_high_ms": 423.07922929824963, "generation_mean_ci95_low_ms": 263.6083773699278, "generation_mean_ms": 343.3438033340887, "generation_min_ms": 274.9132710014237, "generation_p25_ms": 274.9132710014237, "generation_p50_ms": 339.44265700120013, "generation_p75_ms": 415.6754819996422, "generation_p90_ms": 415.6754819996422, "generation_p95_ms": 415.6754819996422, "generation_p99_ms": 415.6754819996422, "generation_samples": 3, "generation_stddev_ms": 70.46214741483334, "mean_inter_token_latency_ms_max": 8.330705181861324, "mean_inter_token_latency_ms_mean": 7.955466337789688, "mean_inter_token_latency_ms_mean_ci95_high": 8.674148700201696, "mean_inter_token_latency_ms_mean_ci95_low": 7.23678397537768, "mean_inter_token_latency_ms_min": 7.222184191514897, "mean_inter_token_latency_ms_p25": 7.222184191514897, "mean_inter_token_latency_ms_p50": 8.313509639992844, "mean_inter_token_latency_ms_p75": 8.330705181861324, "mean_inter_token_latency_ms_p90": 8.330705181861324, "mean_inter_token_latency_ms_p95": 8.330705181861324, "mean_inter_token_latency_ms_p99": 8.330705181861324, "mean_inter_token_latency_ms_samples": 3, "mean_inter_token_latency_ms_stddev": 0.6350991664291974, "mean_time_per_output_token_max_ms": 8.330705181861324, "mean_time_per_output_token_mean_ci95_high_ms": 8.674148700201696, "mean_time_per_output_token_mean_ci95_low_ms": 7.23678397537768, "mean_time_per_output_token_mean_ms": 7.955466337789688, "mean_time_per_output_token_min_ms": 7.222184191514897, "mean_time_per_output_token_p25_ms": 7.222184191514897, "mean_time_per_output_token_p50_ms": 8.313509639992844, "mean_time_per_output_token_p75_ms": 8.330705181861324, "mean_time_per_output_token_p90_ms": 8.330705181861324, "mean_time_per_output_token_p95_ms": 8.330705181861324, "mean_time_per_output_token_p99_ms": 8.330705181861324, "mean_time_per_output_token_samples": 3, "mean_time_per_output_token_stddev_ms": 0.6350991664291974, "output_token_sources": { "usage": 3 }, "output_tokens": 133, "prefix_cache_hit_avg": null, "prompt_token_samples": 3, "prompt_tokens": 913473, "prompt_tokens_max": 304492.0, "prompt_tokens_mean": 304491.0, "prompt_tokens_mean_ci95_high": 304492.1316065276, "prompt_tokens_mean_ci95_low": 304489.8683934724, "prompt_tokens_min": 304490.0, "prompt_tokens_p25": 304490.0, "prompt_tokens_p50": 304491.0, "prompt_tokens_p75": 304492.0, "prompt_tokens_p90": 304492.0, "prompt_tokens_p95": 304492.0, "prompt_tokens_p99": 304492.0, "prompt_tokens_samples": 3, "prompt_tokens_stddev": 1.0, "reasoning_chunks": 0, "request_canary_passed": 0, "request_canary_samples": 0, "throughput_performance_eligible": true, "throughput_tok_s": 0.7631981612491053, "time_to_first_output_max_ms": 61779.706059998716, "time_to_first_output_mean_ci95_high_ms": 61724.81502557574, "time_to_first_output_mean_ci95_low_ms": 53758.90660509143, "time_to_first_output_mean_ms": 57741.86081533359, "time_to_first_output_min_ms": 55322.45940299981, "time_to_first_output_p25_ms": 55322.45940299981, "time_to_first_output_p50_ms": 56123.416983002244, "time_to_first_output_p75_ms": 61779.706059998716, "time_to_first_output_p90_ms": 61779.706059998716, "time_to_first_output_p95_ms": 61779.706059998716, "time_to_first_output_p99_ms": 61779.706059998716, "time_to_first_output_samples": 3, "time_to_first_output_stddev_ms": 3519.7342124284633, "tpot_max_ms": 8.330705181861324, "tpot_mean_ci95_high_ms": 8.674148700201696, "tpot_mean_ci95_low_ms": 7.23678397537768, "tpot_mean_ms": 7.955466337789688, "tpot_min_ms": 7.222184191514897, "tpot_p25_ms": 7.222184191514897, "tpot_p50_ms": 8.313509639992844, "tpot_p75_ms": 8.330705181861324, "tpot_p90_ms": 8.330705181861324, "tpot_p95_ms": 8.330705181861324, "tpot_p99_ms": 8.330705181861324, "tpot_samples": 3, "tpot_stddev_ms": 0.6350991664291974, "ttft_max_ms": 61779.706940000324, "ttft_mean_ci95_high_ms": 61724.815938199434, "ttft_mean_ci95_low_ms": 53758.907045803135, "ttft_mean_ms": 57741.861492001284, "ttft_min_ms": 55322.45968300049, "ttft_p25_ms": 55322.45968300049, "ttft_p50_ms": 56123.41785300305, "ttft_p75_ms": 61779.706940000324, "ttft_p90_ms": 61779.706940000324, "ttft_p95_ms": 61779.706940000324, "ttft_p99_ms": 61779.706940000324, "ttft_samples": 3, "ttft_stddev_ms": 3519.734420942633 }, "model": "glm53-flash-ormandj-sglang-sm120-tp2-393k-c1-adaptive-mtp", "performance_eligible": true, "prompt_cache_mode": "shared", "request_canaries": false, "request_timings": [ { "content_chunks": 17, "controlled_output": null, "decode_tok_s": 138.46226757479639, "e2e_ms": 56462.859640003444, "effective_prefill_tok_s": 5425.364604799794, "generation_ms": 339.44265700120013, "mean_inter_token_latency_ms": 7.222184191514897, "mean_time_per_output_token_ms": 7.222184191514897, "output_token_source": "usage", "output_tokens": 48, "planned_context_tokens": 380000, "prompt_tokens": 304490, "reasoning_chunks": 0, "request_canary": null, "request_index": 0, "time_to_first_output_ms": 56123.416983002244, "tpot_ms": 7.222184191514897, "ttft_ms": 56123.41785300305, "visible_content_truncated": false, "visible_generation_ms": 339.4417870003963 }, { "content_chunks": 16, "controlled_output": null, "decode_tok_s": 120.03785732056966, "e2e_ms": 62054.61933100014, "effective_prefill_tok_s": 4928.657311905739, "generation_ms": 274.9132710014237, "mean_inter_token_latency_ms": 8.330705181861324, "mean_time_per_output_token_ms": 8.330705181861324, "output_token_source": "usage", "output_tokens": 34, "planned_context_tokens": 380000, "prompt_tokens": 304491, "reasoning_chunks": 0, "request_canary": null, "request_index": 1, "time_to_first_output_ms": 61779.706059998716, "tpot_ms": 8.330705181861324, "ttft_ms": 61779.706940000324, "visible_content_truncated": false, "visible_generation_ms": 274.9123909998161 }, { "content_chunks": 24, "controlled_output": null, "decode_tok_s": 120.28614187074675, "e2e_ms": 55738.134884999454, "effective_prefill_tok_s": 5503.949088414699, "generation_ms": 415.6754819996422, "mean_inter_token_latency_ms": 8.313509639992844, "mean_time_per_output_token_ms": 8.313509639992844, "output_token_source": "usage", "output_tokens": 51, "planned_context_tokens": 380000, "prompt_tokens": 304492, "reasoning_chunks": 0, "request_canary": null, "request_index": 2, "time_to_first_output_ms": 55322.45940299981, "tpot_ms": 8.313509639992844, "ttft_ms": 55322.45968300049, "visible_content_truncated": false, "visible_generation_ms": 415.67520199896535 } ], "requests": 3, "response_words": 0, "responses_received": 3, "run_id": "benchmark-20260908T052035Z", "schema": "anvil-serving.benchmark/v1", "serve_flags": { "no_thinking": false, "reasoning_effort": null, "shared_prefix_burst": false, "thinking_mode": "disabled" }, "shared_prefix_tokens": 8000, "source_recipe": null, "started_at": "2026-09-08T05:20:35Z", "started_at_unix_ns": 1788844835453136241, "tier": null, "timestamp_precision": "nanosecond-unix-fields; ISO labels are second precision", "timing_methodology": { "artifact_timestamps": "time.time_ns Unix timestamps; independent artifacts may align them only when the operator supplied the same explicit clock_domain_id", "clock": "client time.perf_counter", "controlled_output": "--response-words adherence is recorded; observe preserves historical behavior while strict rejects non-exact or truncated completions; visible response capture is capped at 8192 characters", "decode": "usage completion tokens after the first token divided by client-observed generation time; completion tokens may include reasoning tokens", "distributions": "request-level mean, sample standard deviation, normal-approximation 95% confidence interval for the mean, min, p25, p50, p75, p90, p95, p99, and max using the nearest-rank percentile method", "effective_prefill": "usage.prompt_tokens divided by client-observed time to first output; includes queueing, scheduling, prefill, and first-token work", "generation": "client-observed E2E minus time to first output", "mean_inter_token_latency": "legacy key for a TPOT proxy: client-observed generation time divided by completion tokens after the first token; not raw per-token ITL", "p99": "numeric nearest-rank p99 is always retained for compatibility; it is descriptive only below 100 eligible samples", "request_canaries": "when enabled, each response must begin with its own synthetic marker and contain no other ANVIL_REQ marker; bounded/truncated captures fail closed and are excluded from performance metrics", "time_to_first_output": "request start through the first non-empty streamed reasoning or content delta", "tpot": "client-observed mean time per output token after the first output; identical to the legacy mean_inter_token_latency key", "ttft": "request start through the first non-empty streamed content delta" }, "wall_clock_ms": 174266.66723400194 }